<p>Group re-identification (GReID) aims to match groups of the same members across different camera perspectives, playing a crucial role in intelligent surveillance systems. A fundamental challenge arises from the intrinsic social affinity within group members, which often leads to densely packed formations and consequently severe mutual occlusions that significantly hinder accurate matching. Although existing GReID models have attempted to tackle occlusion, they still struggle to adapt to the diverse and dynamic nature of real-world occlusions. To address this challenge, we propose a Multi-Granularity Cross-Modal (MGCM) Representation for Occlusion-Invariant GReID. Our approach comprises an Adaptive Occlusion High-granularity(AOH) module and a Complete Feature-based Occlusion Mitigation(CFOM) module. Specifically, AOH efficiently locates the most likely occluded person based on visual geometry priors and the density of pedestrians, leveraging the HRNet pose estimator to identify occluded body parts via keypoint confidence thresholds. To preserve the complete feature dimensions of these identified regions, the CFOM module integrates the CLIP model to generate multi-dimensional semantic descriptions, which are then fused with visual features through a dynamic weight adjustment mechanism. This design effectively mitigates occlusion-induced noise while retaining discriminative information, thus achieving an optimal balance between suppressing false features and preventing feature loss. Extensive experiments on public datasets demonstrate that our method achieves state-of-the-art performance, delivering significant improvements in the robustness and accuracy of GReID in challenging occlusion scenarios.</p>